NLM DIR Seminar Schedule

Abstract:

Advances in multimodal AI have opened new opportunities for automated lesion analysis in computed tomography (CT), yet progress is limited by the lack of richly annotated datasets and robust evaluation benchmarks. In this seminar, I will present CT-Bench, a large-scale, clinically grounded CT dataset containing 20,335 lesions with bounding boxes, radiologist-derived descriptions, and size measurements, along with a comprehensive QA benchmark covering seven key lesion analysis tasks such as localization, image retrieval, description generation, and attribute classification. I will also introduce CARE-CT, a consistency-aware reflective agent designed to perform multi-step CT reasoning with enhanced reliability. CARE-CT integrates domain-specific tools with consistency scoring and adaptive reflection to identify and correct reasoning errors. Together, CT-Bench and CARE-CT provide a strong foundation for training, evaluating, and improving multimodal AI systems for real-world CT interpretation.